fix(ai): align Anthropic and OpenAI message conversion - #1949

Summary

• Normalize Anthropic user and assistant turns as content blocks, then merge consecutive compatible turns without flattening their blocks.
• Preserve explicit empty assistant content for OpenAI tool-call replay, while omitting the field only when it is absent.
• Stop sending name on OpenAI tool messages.

Type and Areas

Type:

Bug fix

Areas:

AI adapters

Motivation / Impact

Keeps provider replay payloads faithful to their source message shape: Anthropic consecutive turns retain block boundaries, and OpenAI-compatible tool-call turns no longer rewrite "" as whitespace or attach a tool-message name.

Verification

• pnpm run fmt:rs
• cargo test -p bitfun-ai-adapters message_converter — 26 passed
• git diff --cached --check
• cargo check --workspace

Reviewer Notes

• The existing OpenAI Chat Completions tool-image behavior is intentionally unchanged in this PR.
• Scope is limited to the two provider message converters.

Checklist

• This PR is focused and does not include secrets, temporary prompts, generated scratch files, or unrelated artifacts.
• Relevant verification is recorded above, including the incomplete workspace check.
• No user-facing strings, docs, or locales were changed.